Agents can improve their tools by grouping shared abilities
The Trinity:Capability-Driven Recursive Self-Improvement for Multimodal Agents · arXiv:2610.36077
In plain words
The Trinity is a framework for improving the software that organizes a multimodal agent’s actions and reusable modules, while keeping the underlying model fixed.
It groups tasks that need similar abilities, creates simpler practice worlds for those abilities, and uses feedback to improve the software.
Across averages for 28 training tasks and 18 unseen tasks, the fourth version scored higher than the starting version. The Base Harness was evaluated without learned modules or prepared skills; each evolved version used skills prepared from five trials for that Harness.
The comparison therefore covers both the evolved software and the skills used with evolved versions.
The research
As agents face varied tasks, building checkable practice environments is costly, while tuning one action system for each task may limit transfer. Can shared abilities provide a way to organize improvement that carries across tasks? The Trinity tests that question by grouping tasks around the capabilities they require.
The Trinity groups tasks by shared abilities, builds verifiable practice worlds, and evolves reusable action modules while keeping the underlying model fixed.
To test transfer, the team compared capability trees after one round: one learned on original tasks, one on verified simulated worlds. Both used the same tree structure and module-only evaluation.
Task drivers scored outcomes; researchers averaged task-specific 0–100 scores equally across 28 training and 18 unseen tests. Simulated-world training scored higher on both splits, supporting transfer, though token costs differed.
The score comparisons are averages across tasks, and results vary by task. The framework changes the Harness while keeping the model’s underlying weights fixed; the embodied comparison also includes separately prepared skills, so it does not isolate module evolution from skill preparation.
The authors say web-search coverage and result quality may vary, and the study focuses mainly on Harness-level improvement.
What the evidence shows
- Improve across varied tasks
The Trinity is a recursive self-improvement framework for multimodal agents. It works on a Harness, the software that organizes an agent’s actions and reusable modules, rather than changing the underlying model.
The paper’s question is how to improve this software across varied tasks when building verifiable practice environments is costly and tuning a separate system for each task may limit transfer.
- Organize around shared abilities
The framework groups tasks according to abilities they share, then builds a hierarchy called a capability tree. Creator makes simpler, verifiable practice worlds for the shared requirements; Explorer refines reusable Harness modules using feedback from those worlds; and a router selects a compatible Harness for a task.
Evaluating the resulting configurations on original tasks tests whether practice transfers beyond the generated worlds.
- Measure repeated improvement
The authors track the starting Harness through four improvement rounds. Task drivers score each task, and researchers average the task-specific scores equally within the training and unseen splits.
The means rise from 39.06 to 73.16 points across 28 training tests and from 37.45 to 63.61 across 18 unseen tests. The Base Harness has no learned modules or skills; each evolved Harness version uses skills prepared specifically for it. The comparison therefore reflects both Harness evolution and the skills used with evolved versions.
- Compare simulated practice
To check whether the practice worlds support transfer, the authors compare two capability trees after one round. Both use the same tree structure and are evaluated with modules only, without skills; one learns from original tasks and the other from verified simulated worlds.
The simulated version has higher mean scores on both splits, although it uses more training tokens, so the comparison does not isolate curriculum alone.
- Test selection and routing
A module-addition comparison tests whether task relevance matters: adding a relevant module to the Base Harness produces a higher training mean than adding an irrelevant one. The routing comparison then tests whether an agent can select suitable Harnesses using task information and verified experience.
Its average scores come close to a retrospective Oracle reference, which can choose the best compatible Harness after seeing each instance and is not an available online method.
- Extend evidence to robots
The study also evaluates complete Harness-and-skill configurations on two physical robot tasks. The Base configuration succeeds in none of the test episodes on either task, while the fourth iteration succeeds in 60% on each.
Because every Harness has separately prepared skills, these outcomes extend the evidence to physical tasks but cannot attribute the improvement to module evolution alone.
Deep
- Original paper authors
- Weidong Tang, Ruiheng Wang, Jingxuan Zhang, Hexin Wang, Haoyuan Chen
- Submitted to arXiv
- License
- CC BY 4.0